Safe and accelerated screening framework for support tensor machines.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Support Tensor Machines (STMs) are effective for high-dimensional tensor data classification.
• Traditional iterative methods for STMs are often computationally intensive and time-consuming.
• Accelerating STM training is crucial for practical applications.

Purpose of the Study:

• To develop a novel safe screening rule to accelerate Support Tensor Machines (STMs) training.
• To reduce the computational complexity and training time for high-dimensional tensor data classification.
• To ensure the safety and effectiveness of the proposed screening methods.

Main Methods:

• Generalizing safe screening strategies from Support Vector Machines (SVMs) to the tensor domain.
• Proposing a dual static screening rule (DSSR) using variational inequalities to pre-screen samples.
• Introducing a dynamic screening rule (DGSR) utilizing the duality gap for iterative sample screening during training.
• Developing a flexible DS-DGSR framework integrating DSSR and DGSR with a post-training verification step based on optimality conditions.

Main Results:

• The proposed screening rules effectively reduce the problem scale and accelerate STM training.
• The DS-DGSR framework demonstrates flexibility in handling various tensor decomposition methods and data characteristics.
• Numerical experiments on real-world high-dimensional tensor datasets validate the effectiveness and feasibility of the DS-DGSR framework.
• The screening process significantly reduces training time without compromising classification accuracy.

Conclusions:

• The novel safe screening rule and DS-DGSR framework offer a significant improvement in the efficiency of Support Tensor Machines.
• This approach provides a practical solution for accelerating the classification of large-scale tensor data.
• The method is adaptable and robust, applicable to diverse STM implementations and datasets.